stats 700-002 data analysis using pythonpie chart; the only worse design than a pie chart is several...
TRANSCRIPT
STATS 700-002Data Analysis using Python
Lecture 5: numpy and matplotlibSome examples adapted from A. Tewari
Reminder!If you don’t already have a Flux/Fladoop username, request one promptly!
Make sure you have a way to ssh to the Flux clusterUNIX/Linux/MacOS: you’re all set!Windows:
install PuTTY:https://www.chiark.greenend.org.uk/~sgtatham/putty/latest.html
and you may also want cygwin https://www.cygwin.com/
You also probably want to set up VPN to access Flux from off-campus:http://its.umich.edu/enterprise/wifi-networks/vpn
Numerical computing in Python: numpyOne of a few increasingly-popular, free competitors to MATLAB
Numpy quickstart guide: https://docs.scipy.org/doc/numpy-dev/user/quickstart.html
For MATLAB fans:https://docs.scipy.org/doc/numpy-dev/user/numpy-for-matlab-users.html
Closely related package scipy is for optimizationSee https://docs.scipy.org/doc/
numpy data types:Five basic numerical data types:
boolean (bool)integer (int)unsigned integer (uint)floating point (float)complex (complex)
Many more complicated data types are availablee.g., each of the numerical types can vary in how many bits it useshttps://docs.scipy.org/doc/numpy/user/basics.types.html
numpy.array: numpy’s version of Python array (i.e., list)Can be created from a Python list…
...by “shaping” an array…
...by “ranges”...
...or reading directly from a filesee https://docs.scipy.org/doc/numpy/user/basics.creation.html
numpy allows arrays of arbitrary dimension (tensors)1-dimensional arrays:
2-dimensional arrays (matrices):
3-dimensional arrays (“3-tensor”):
More on numpy.arange creationnp.arange(x): array version of Python’s range(x), like [0,1,2,...,x-1]
np.arange(x,y): array version of range(x,y), like [x,x+1,...,y-1]
np.arange(x,y,z): array of elements [x,y) in z-size increments.
Related useful functions, that give better/clearer control of start/endpoints and allow for multidimensional arrays:
https://docs.scipy.org/doc/numpy/reference/generated/numpy.linspace.htmlhttps://docs.scipy.org/doc/numpy/reference/generated/numpy.ogrid.htmlhttps://docs.scipy.org/doc/numpy/reference/generated/numpy.mgrid.html
More on numpy.arange creationnp.arange(x): array version of Python’s range(x), like [0,1,2,...,x-1]
np.arange(x,y): array version of range(x,y), like [x,x+1,...,y-1]
np.arange(x,y,z): array of elements [x,y) in z-size increments.
numpy array indexing is highly expressive...
Not very relevant to us right now…
...but this will come up again in a few weeks when we cover TensorFlow!
More array indexingNumpy allows MATLAB/R-like indexing by Booleans
This error is by design, believe it or not! The designers of numpy were concerned about ambiguities in Boolean vector operations, so they split the two operations into two separate methods, x.any() and x.all()
Boolean operations: np.any(), np.all()Analogous to and and or, respectively
axis argument picks which axis along which to perform the Boolean operation. If left unspecified, it treats the array as a single vector.
Setting axis to be the first (i.e., 0-th) axis yields the entrywise behavior we wanted.
Boolean operations: np.logical_and()Numpy also has built-in Boolean vector operations, which are simpler/clearer at the cost of the expressiveness of np.any(), np.all().
This is an example of a numpy “universal function” (ufunc), which we’ll discuss more in a few slides.
Random numbers in numpynp.random contains methods for generating random numbers
Lots more distributions:https://docs.scipy.org/doc/numpy/reference/routines.random.html#distributions
np.random.choice(): random samples from datanp.random.choice(x,[size,replace,p])
Generates a sample of size elements from the array x, drawn with (replace=True) or without (replace=False) replacement, with element probabilities given by vector p.
shuffle() vs permutation()np.random.shuffle(x)
randomly permutes entries of x in placeso x itself is changed by this operation!
np.random.permutation(x)returns a random permutation of xand x remains unchanged.
Statistics in numpyNumpy implements all the standard statistics functions you’ve come to expect
Statistics in numpy (cont’d)Numpy deals with NaNs more gracefully than MATLAB/R:
For more basic statistical functions, see: https://docs.scipy.org/doc/numpy-1.8.1/reference/routines.statistics.html
Probability and statistics in scipyAll the distributions you could possibly ever want:
https://docs.scipy.org/doc/scipy/reference/stats.html#continuous-distributionshttps://docs.scipy.org/doc/scipy/reference/stats.html#multivariate-distributionshttps://docs.scipy.org/doc/scipy/reference/stats.html#discrete-distributions
More statistical functions (moments, kurtosis, statistical tests):https://docs.scipy.org/doc/scipy/reference/stats.html#statistical-functions
Second argument is the name of a distribution in scipy.stats
Kolmogorov-Smirnov test
numpy/scipy universal functions (ufuncs)From the documentation:
A universal function (or ufunc for short) is a function that operates on ndarrays in an element-by-element fashion, supporting array broadcasting, type casting, and several other standard features. That is, a ufunc is a “vectorized” wrapper for a function that takes a fixed number of scalar inputs and produces a fixed number of scalar outputs.https://docs.scipy.org/doc/numpy/reference/ufuncs.html
So ufuncs are vectorized operations, just like in R and MATLAB
ufuncs in actionlist comprehensions are great, but they’re not well-suited to numerical computing
Sorting with numpy/scipy
Sorting is along the “last” axis by default. Note contrast with np.any() . To treat the array as a single vector, axis must be set to None.
ASCII rears its head-- capital letters are “earlier” than all lower-case by default.
Original array is unchanged by use of np.sort() , like Python’s built-in sorted()
A cautionary notenumpy/scipy have a number of similarly-named functions with different behaviors!
Example: np.amax, np.ndarray.max, np.maximum
The best way to avoid these confusions is to1) Read the documentation carefully2) Test your code!
Plotting with matplotlibmatplotlib is a plotting library for use in Python
Similar to R’s ggplot2 and MATLAB’s plotting functions
For MATLAB fans, matplotlib.pyplot implements MATLAB-like plotting:http://matplotlib.org/users/pyplot_tutorial.html
Sample plots with code:http://matplotlib.org/tutorials/introductory/sample_plots.html
Basic plotting: matplotlib.pyplot.plotmatplotlib.pyplot.plot(x,y)
plots y as a function of x.
matplotlib.pyplot(t)sets x-axis to np.arange(len(t))
Basic plotting: matplotlib.pyplot.plot
Jupyter “magic” command to make images appear in-line.
Python ‘_’ is a placeholder, similar to MATLAB ‘~’. Tells Python to treat this like variable assignment, but don’t store result anywhere.
Customizing plotsSecond argument to pyplot.plot specifies line type, line color, and marker type. Specify broader array of colors, line weights, markers, etc., using long-hand arguments.
Customizing plots
Long form of the command on the previous slide. Same plot!
A full list of the long-form arguments available to pyplot.plot are available in the table titled“Here are the available Line2D properties.”:http://matplotlib.org/users/pyplot_tutorial.html
Multiple lines in a single plot
Note: more complicated specification of individual lines can be achieved by adding them to the plot one at a time.
Multiple lines in a single plot: long form
Note: same plot as previous slide, but specifying one line at a time so we could, if we wanted, use more complicated line attributes.
Titles and axis labels
Titles and axis labelsChange font sizes
Legends
pyplot.legend generates legend based on label arguments passed to pyplot.plot . loc=‘best’ tells pyplot to place the legend where it thinks is best.
Can use LaTeX in labels, titles, etc.
Gridlines on/off: plt.grid
Annotating figures
Specify text coordinates and coordinates of the arrowhead using the coordinates of the plot itself. This is pleasantly different from many other plotting packages, which require specifying coordinates in pixels!
Plotting histograms: pyplot.hist()
Plotting histograms: pyplot.hist()
Bin counts. Note that if normed=1 , then these will be proportions between 0 and 1 instead of counts.
Bar plotsbar(x, height, *, align='center', **kwargs)
Full set of available arguments to bar(...) can be found at http://matplotlib.org/api/_as_gen/matplotlib.pyplot.bar.html#matplotlib.pyplot.bar
Horizontal analogue given by barh http://matplotlib.org/api/_as_gen/matplotlib.pyplot.barh.html#matplotlib.pyplot.barh
Tick labels
Can specify what the x-axis tic labels should be by using the tick_label argument to plot functions.
Box & whisker plots
plt.boxplot(x,...) : x is the data. Many more optional arguments are available, most to do with how to compute medians, confidence intervals, whiskers, etc. See http://matplotlib.org/api/_as_gen/matplotlib.pyplot.boxplot.html#matplotlib.pyplot.boxplot
Don’t use pie charts!
But if you must…pyplot.pie(x, … )http://matplotlib.org/api/_as_gen/matplotlib.pyplot.pie.html#matplotlib.pyplot.pie
Pie Charts
A table is nearly always better than a dumb pie chart; the only worse design than a pie chart is several of them, for then the viewer is asked to compare quantities located in spatial disarray both within and between charts [...] Given their low [information] density and failure to order numbers along a visual dimension, pie charts should never be used.
Edward TufteThe Visual Display of Quantitative Information
Subplotssubplot(nrows, ncols, plot_number)
Shorthand: subplot(XYZ)Makes an X-by-Y plotPicks out the Z-th plotCounting in row-major order
tight_layout() automatically tries to clean things up so that subplots don’t overlap. Without this command in this example, the labels “sqrt” and “logarithmic” overlap with the x-axis tick labels in the first row.
Specifying axis rangesplt.ylim([lower,upper]) sets y-axis limits
plt.xlim([lower,upper]) for x-axis
For-loop goes through all of the subplots and sets their y-axis limits
Nonlinear axesScale the axes with plt.xscale and plt.yscale
Built-in scales:Linear (‘linear’)Log (‘log’)Symmetric log (‘symlog’)Logit (‘logit’)
Can also specify customized scales: https://matplotlib.org/devel/add_new_projection.html#adding-new-scales
Saving imagesplt.savefig(filename) will try to automatically figure out what file type you want based on the file extension.
Can make it explicit using plt.savefig(‘filename’,
format=‘fmt’)
Other options for specifying resolution, padding, etc: https://matplotlib.org/api/_as_gen/matplotlib.pyplot.savefig.html
AnimationsMatplotlib.animate package generates animations
I won’t require you to make any, but they’re fun to play around with (and they can be a great visualization tool)
The details are a bit tricky, so I recommend starting by looking at some of the example animations here: http://matplotlib.org/api/animation_api.html#examples
Readings (this lecture)Required:
Numpy quickstart tutorial:https://docs.scipy.org/doc/numpy-dev/user/quickstart.html
Pyplot tutorial:http://matplotlib.org/tutorials/introductory/pyplot.html#sphx-glr-tutorials-introductory-pyplot-py
Recommended:The Visual Display of Quantitative Information by Edward TufteVisual and Statistical Thinking: Displays of Evidence for Making Decisions
by Edward TufteThis is a short book, essentially a reprint of Chapter 2 of the book above.
Pyplot API: http://matplotlib.org/api/pyplot_summary.html
Readings (next lecture)Required:
Introduction to Unix commands: https://kb.iu.edu/d/afskIncludes all the commands we discussed today, and a few more that you don’t need to know well, but are worth being aware of.
Recommended:Survival guide for Unix newbies: http://matt.might.net/articles/basic-unix/
More thorough discussion, including advanced commands like grep
“GNU/Linux Command−Line Tools Summary” by Gareth AndersonComprehensive introduction to the command line and the UNIX/Linux design philosophy in general.http://tldp.org/LDP/GNU-Linux-Tools-Summary/GNU-Linux-Tools-Summary.pdf